The one to remember: automating a broken process. AI runs whatever you give it faster and at scale — so a flawed workflow just produces flawed output more quickly. Fix the process first.
The ten mistakes
- Automating a broken process. AI scales whatever you give it. Fix the workflow before you automate it, or you industrialise the flaw.
- Skipping data-quality checks. Messy inputs produce garbage outputs. If you wouldn't trust a new hire to act on this data, don't trust AI with it either.
- No human review step. Treating AI output as final on anything that matters. Keep a human in the loop where consequences are real.
- Defaulting to the most expensive model. Most tasks run fine on a mid-tier or budget model. Paying frontier prices for classification is pure waste — match the model to the task.
- Starting too broad. Trying to "do AI" everywhere at once. Pick one workflow, prove it, then expand — see the starter guide.
- Not version-locking prompts. Prompts that change silently break reproducibility. Treat system prompts like code: version them, test them, log changes.
- Trusting outputs without verification. Confident does not mean correct. Models state fabrications with full confidence — see the Truth Score.
- Ignoring the agentic cost multiplier. Agents use 5–20x more tokens than a single completion. Model real cost in the calculator before deploying agents.
- Putting sensitive data into non-compliant tools. Check residency and compliance first — see the privacy checklist.
- No owner, no metric. Projects with no single owner and no measurable goal drift and quietly die. Every AI initiative needs both.
The pattern behind the failures
Notice the theme: the failures are about process and discipline, not the technology. The model is rarely the problem. The companies getting value are not using better AI — they are starting narrow, checking outputs, matching cost to task, and measuring results.
Quick self-audit
| Ask yourself | If no… |
|---|---|
| Is the process we're automating actually working today? | Fix it first |
| Does a human review anything consequential? | Add a review step |
| Are we on the cheapest model that does the job? | Re-check model choice |
| Does this project have an owner and a metric? | Assign both |
| Have we checked data compliance? | Run the privacy checklist |
What changed in June 2026
- As adoption hit ~42% of small businesses, the gap widened between disciplined adopters and those who rushed in.
- The agentic cost multiplier became the most common budget surprise as agents went mainstream.
- Prompt versioning emerged as a real operational practice, not just a developer nicety.
Avoiding the traps? Start with the 4-phase starter guide and know what AI can't do.